Variational quantum circuit training method and apparatus, electronic device, and medium

Abstract

This invention provides a variable quantum circuit training method, apparatus, electronic device, and medium. The method, used to provide a quantum-enhanced feature extraction module for motion artifact detection in blood pressure signals, includes the following steps: acquiring synchronously collected PPG and motion reference signals, preprocessing them to construct labeled data segments; extracting static and temporal feature matrices for each segment; inputting the static features into a quantum module to generate quantum-enhanced feature vectors; fusing the quantum and temporal features to predict the signal quality level using a CNN; combining classification loss and differential consistency constraint loss to construct an overall loss; updating model parameters with the goal of minimizing the overall loss to complete training. This method utilizes geometric and physical constraints to enhance features, achieving accurate artifact detection in various motion scenarios. It is highly accurate, lightweight, and suitable for wearable blood pressure monitoring devices.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing and wearable device technology, specifically relating to a variable quantum circuit training method, device, electronic device and medium. Background Technology

[0002] In wearable devices using photoplethysmography (PPG) for cuffless continuous blood pressure monitoring, strong motion artifacts are introduced into the PPG signal due to human movement. These motion artifacts exhibit a deeply nonlinear coupling relationship with the physiological pulse wave component, and their frequency bands highly overlap with the effective signal. Therefore, traditional linear filtering, adaptive filtering, or feature extraction methods based on fixed thresholds are insufficient to effectively separate or accurately identify motion artifacts, resulting in inadequate accuracy and reliability in signal quality assessment and severely limiting the clinical applicability of blood pressure monitoring in dynamic scenarios. Summary of the Invention

[0003] In view of this, in order to overcome at least one aspect of the above-mentioned problems, the present invention proposes a variable quantum circuit training method for providing a quantum-enhanced feature extraction module for motion artifact detection of blood pressure signals, comprising the following steps: Obtain a training dataset containing multiple multi-channel signal segments. Each multi-channel signal segment is obtained by preprocessing and window segmentation of the synchronously acquired raw PPG signal and raw motion reference signal, and each multi-channel signal segment corresponds to a real label. Construct the static feature vector and temporal feature matrix for each of the multi-channel signal segments; The static feature vector is input into the quantum feature extraction module for quantum feature extraction processing to obtain the quantum enhanced feature vector; The quantum-enhanced feature vector and the temporal feature matrix are fused together, and the fused features are input into a convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. Based on the change intensity of the quantum enhancement eigenvector corresponding to adjacent time windows and the standard deviation of the original PPG signal differential sequence within the corresponding window, the differential consistency constraint loss is calculated; wherein, the differential consistency constraint loss is used to ensure that the transformation characteristics of the quantum enhancement eigenvector relative to the static eigenvector are consistent with the differential characteristics of the original PPG signal and the original motion reference signal; The classification loss is calculated based on the probability prediction results and the true labels, and the classification loss is weighted and summed with the differential consistency constraint loss to obtain the total loss. With minimizing the overall loss as the optimization objective, the trainable parameters of the quantum feature extraction module and the convolutional neural network classifier are iteratively updated using the gradient descent method to complete the training.

[0004] Furthermore, in the step of obtaining the training dataset, the preprocessing and window segmentation include the following steps: The original PPG signal and the original motion reference signal were synchronously sampled using the same master clock, and the sampling frequency was set to be no less than 100 Hz. The original PPG signal is subjected to a first bandpass filter with a passband frequency range of 0.5 Hz to 8.0 Hz; The original motion reference signal is subjected to a second bandpass filter with a passband frequency range of 0.1 Hz to 20.0 Hz; The filtered PPG signal and the motion reference signal are windowed to obtain the original multi-channel signal segments. The original multi-channel signal segment is normalized to eliminate signal amplitude scaling and DC offset caused by individual and equipment differences, thus obtaining the multi-channel signal segment.

[0005] Further, constructing the static feature vector and temporal feature matrix of the multi-channel signal segment includes the following steps: The first-order difference and the second-order difference are calculated for the PPG signal and the motion reference signal in the multi-channel signal segment, respectively, to obtain the differential sequence. Based on the differential sequence, a first set of scalar features is extracted. The first set of scalar features includes at least one of standard deviation, kurtosis and zero-crossing rate. The absolute integral and square integral within the sliding window are calculated for the PPG signal and motion reference signal in the multi-channel signal segment, respectively, to obtain an integral sequence. Based on the integral sequence, a second set of scalar features is extracted. The second set of scalar features includes at least one of mean, coefficient of variation, and root mean square. Calculate the normalized cross-correlation function between the differential sequence of the PPG signal and the differential sequence of the motion reference signal, extract the maximum value of the cross-correlation coefficient and its corresponding time delay, and calculate the energy ratio of the PPG signal and the motion reference signal, which are used as the third set of scalar features. The first set of scalar features, the second set of scalar features, and the third set of scalar features are concatenated into a static feature vector, and one or more of the original PPG signal, the original motion reference signal, and its first-order differential sequence are retained and arranged in time sequence to form a time-series feature matrix.

[0006] Furthermore, the quantum feature extraction module includes a classical encoder, an angle encoding layer, a variable quantum circuit, a measurement layer, and a linear layer.

[0007] Further, the step of inputting the static feature vector into the quantum feature extraction module for quantum feature extraction processing to obtain the quantum enhanced feature vector includes: The static feature vector is mapped to a rotation angle parameter using the classic encoder. The rotation angle parameter is loaded onto the rotation gate of the qubit through the angle encoding layer to generate the input quantum state of the quantum circuit. The input quantum state is transformed using the parametric quantum gates and entanglement gates in the variable quantum circuit to obtain the final quantum state; The final quantum state is measured through the measurement layer, and the measurement result is linearly transformed through the linear layer to obtain the quantum enhancement feature vector.

[0008] Further, the feature fusion of the quantum-enhanced feature vector and the temporal feature matrix includes: The quantum enhancement feature vector is copied along the time dimension to generate a quantum feature matrix with the same time length and identical rows as the temporal feature matrix; The temporal feature matrix and the quantum feature matrix are concatenated along the feature channel dimension to obtain a fused feature matrix; The fused feature matrix is ​​converted into a three-dimensional tensor.

[0009] Furthermore, the convolutional neural network is composed of depthwise separable convolutional layers, dilated convolutional layers, global average pooling layers, and at least one fully connected layer connected in series.

[0010] Furthermore, the gradient descent method employs the Adam optimizer to perform parameter iteration; for the trainable parameters of the variable quantum circuit, the gradient is calculated using parameter shift rules or automatic differentiation.

[0011] A second aspect of the present invention provides a variable quantum circuit training device, comprising: The data acquisition module is used to acquire a training dataset, which contains multiple multi-channel signal segments. Each multi-channel signal segment is obtained by preprocessing and window segmentation of the synchronously acquired raw PPG signal and raw motion reference signal, and each multi-channel signal segment corresponds to a real label. The feature construction module is used to construct the static feature vector and temporal feature matrix for each of the multi-channel signal segments; A quantum feature extraction module is used to perform quantum feature extraction processing on the static feature vector to obtain a quantum enhanced feature vector; The fusion classification module is used to fuse the quantum-enhanced feature vector with the temporal feature matrix, and input the fused features into a convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. The loss calculation module is used to calculate the differential consistency constraint loss based on the change intensity of the quantum enhancement feature vector corresponding to adjacent time windows and the standard deviation of the original PPG signal difference sequence within the corresponding window, calculate the classification loss based on the probability prediction result and the true label, and perform a weighted summation of the classification loss and the differential consistency constraint loss to obtain the total loss. The parameter update module is used to iteratively update the trainable parameters of the quantum feature extraction module and the convolutional neural network classifier using gradient descent, with the goal of minimizing the overall loss, to complete the training.

[0012] According to a third aspect of the invention, a non-transitory computer-readable storage medium is provided storing computer instructions for causing a computer to perform any of the methods in the first aspect.

[0013] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements any of the methods in the first aspect.

[0014] In a fifth aspect, the present invention proposes a method for detecting motion artifacts in blood pressure signals, based on a quantum feature extraction module and a convolutional neural network classifier trained using the training method described in any of the above embodiments, comprising the following steps: The original PPG signal and the original motion reference signal are acquired synchronously, preprocessed and windowed to obtain the multi-channel signal segment to be detected; Construct the static feature vector and temporal feature matrix of the multi-channel signal segment to be detected; The static feature vector is input into the trained quantum feature extraction module to obtain the quantum enhanced feature vector; The quantum-enhanced feature vector is fused with the temporal feature matrix, and the fused features are input into the trained convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. Based on the probability prediction results, a hierarchical decision is executed. The hierarchical decision includes at least adopting the PPG signal of the current window for blood pressure calculation, adopting it after marking it for blood pressure calculation, or rejecting it for blood pressure calculation, and triggering user feedback when rejecting it.

[0015] Furthermore, the probability prediction result includes a first quality level, a second quality level, and a third quality level, and the step of performing a hierarchical decision based on the probability prediction result includes: Set a high confidence threshold and a low confidence threshold, wherein the high confidence threshold is greater than the low confidence threshold; The category corresponding to the maximum value among the probabilities of the first quality level, the second quality level, and the third quality level is taken as the predicted main category. When the predicted main category is the first quality level and its probability is greater than or equal to the high confidence threshold, the decision is to adopt the current signal. When the predicted main category is the first quality level and its probability is between the low confidence threshold and the high confidence threshold, or when the predicted main category is the second quality level, or when the predicted main category is the third quality level and its probability is less than the low confidence threshold, the decision is to label and then adopt. When the predicted main category is the third quality level and its probability is greater than or equal to the high confidence threshold, the decision is to reject the current signal.

[0016] Furthermore, when the decision is to adopt the current signal, the PPG signal of the current window is passed to the blood pressure calculation algorithm, and the calculated blood pressure value is marked as a high-confidence measurement result.

[0017] Furthermore, when the decision is to adopt the marked value, the PPG signal of the current window is passed to the blood pressure calculation algorithm, and the calculated blood pressure value is marked as a low-confidence measurement result.

[0018] Furthermore, when the decision is to reject the current signal, the PPG signal of the current window is discarded, the blood pressure calculation process is not triggered, and the user feedback mechanism is activated.

[0019] Furthermore, the triggering of user feedback includes: Obtain the user's current state, which includes both active and inactive states; When the decision is to reject the current signal and the user is active, a vibration alert is triggered. A visual cue is triggered when the decision is to reject the current signal and the user is stationary.

[0020] Furthermore, it also includes: after triggering a vibration alert or visual alert, the system enters a silent period of preset duration, during which user feedback is prohibited from being triggered again.

[0021] In a sixth aspect, the present invention provides a blood pressure signal motion artifact detection device, based on the blood pressure signal motion artifact detection method described in any of the above embodiments, the device comprising: The signal acquisition and preprocessing module is used to simultaneously acquire PPG and motion reference signals, and output multi-channel signal segments to be detected after preprocessing and window segmentation. The feature construction module is used to construct static feature vectors and temporal feature matrices based on the multi-channel signal segments to be detected. The quantum feature extraction module is configured to take a static feature vector as input and output a quantum-enhanced feature vector. The feature fusion classification module is used to fuse features and output quality level probability prediction results based on the trained convolutional neural network classifier. The decision feedback module is used to perform hierarchical decisions based on probability prediction results. The hierarchical decisions include at least adopting the PPG signal of the current window for blood pressure calculation, adopting it after marking it for blood pressure calculation, or rejecting it for blood pressure calculation, and triggering user feedback when rejecting it.

[0022] According to a seventh aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any method of the first aspect, or to perform any method of the fifth aspect.

[0023] According to an eighth aspect of the invention, a non-transitory computer-readable storage medium is provided storing computer instructions for causing a computer to perform any of the methods in the first aspect or any of the methods in the fifth aspect.

[0024] According to a ninth aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements any of the methods in the first aspect or any of the methods in the fifth aspect.

[0025] This training method utilizes mathematical geometric features to characterize the dynamic nature of signals and enhances feature separability through quantum transformations constrained by physical laws. Ultimately, it achieves accurate and robust detection of motion artifacts in blood pressure signals across various scenarios, from static to strenuous exercise. It significantly improves signal quality detection accuracy and generalization ability under motion interference. The algorithm's lightweight design makes it suitable for embedded deployment, providing crucial technical support for reliable, continuous, and non-invasive blood pressure monitoring in wearable devices. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0027] Figure 1 is a flowchart illustrating the variable quantum circuit training method in an embodiment of the present invention; Figure 2 is a schematic diagram of the preprocessing and window segmentation process in an embodiment of the present invention; Figure 3 is a flowchart illustrating the construction of static feature vectors and temporal feature matrices in an embodiment of the present invention; Figure 4 is a schematic diagram of the quantum feature extraction process in an embodiment of the present invention; Figure 5 is a schematic diagram of the feature fusion process in an embodiment of the present invention; Figure 6 is a structural block diagram of the variable quantum circuit training device in an embodiment of the present invention; Figure 7 is a flowchart illustrating the method for detecting motion artifacts in blood pressure signals in an embodiment of the present invention; Figure 8 is a schematic diagram of the hierarchical decision-making process in an embodiment of the present invention; Figure 9 is a structural block diagram of the blood pressure signal motion artifact detection device in an embodiment of the present invention; Figure 10 is a structural block diagram of the electronic device in an embodiment of the present invention.

[0028] Figure Labels Variable quantum circuit training device 100; data acquisition module 110; feature construction module 120; quantum feature extraction module 130; fusion classification module 140; loss calculation module 150; parameter update module 160; Blood pressure signal motion artifact detection device 200; signal acquisition and preprocessing module 210; feature construction module 220; quantum feature extraction module 230; feature fusion and classification module 240; decision feedback module 250; Electronic device 300; computing unit 310; ROM 320; RAM 330; bus 340; I / O interface 350; input unit 360; output unit 370; storage unit 380; communication unit 390. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to specific examples and the accompanying drawings.

[0030] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities with the same name but different names or different parameters. It can be seen that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of the present invention. Subsequent embodiments will not explain this in detail.

[0031] According to an embodiment of the present invention, a variable quantum circuit training method is used to provide a quantum-enhanced feature extraction module for motion artifact detection of blood pressure signals, such as... Figure 1 As shown, the method includes the following steps: Step S10: Obtain the training dataset, which contains multiple multi-channel signal segments. Each multi-channel signal segment is obtained by preprocessing and window segmentation of the synchronously acquired raw PPG signal and raw motion reference signal, and each multi-channel signal segment corresponds to a real label. It's important to note that the goal of this step is to transform the raw, noisy, continuous signals acquired from wearable device sensors into a series of clean, standardized, time-aligned data segments (windows) that are easy for subsequent deep processing. This is the foundation of the entire process, and its quality directly determines the success or failure of subsequent advanced feature extraction and classification.

[0032] Step S20: Construct the static feature vector and temporal feature matrix for each of the multi-channel signal segments; Step S30: Input the static feature vector into the quantum feature extraction module for quantum feature extraction processing to obtain the quantum enhanced feature vector; Step S40: Perform feature fusion between the quantum enhanced feature vector and the temporal feature matrix, and input the fused features into a convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window; Step S50: Calculate the differential consistency constraint loss based on the change intensity of the quantum enhancement feature vector corresponding to adjacent time windows and the standard deviation of the original PPG signal differential sequence within the corresponding window; wherein, the differential consistency constraint loss is used to ensure that the transformation characteristics of the quantum enhancement feature vector relative to the static feature vector are consistent with the differential characteristics of the original PPG signal and the original motion reference signal. Step S60: Calculate the classification loss based on the probability prediction result and the true label, and then perform a weighted summation of the classification loss and the differential consistency constraint loss to obtain the total loss; Step S70: With minimizing the overall loss as the optimization objective, the trainable parameters of the quantum feature extraction module and the convolutional neural network classifier are iteratively updated using the gradient descent method to complete the training.

[0033] In other words, firstly, the synchronously acquired PPG and accelerometer signals are preprocessed and windowed. Secondly, the differential, integral, and cross-differential features of the signals are calculated in parallel to construct static feature vectors and temporal feature matrices. Then, the static feature vectors are input into a variable quantum circuit via a trainable encoder to perform feature enhancement transformation in quantum Hilbert space, and differential consistency constraints are introduced during training to ensure that the changes in quantum features are consistent with the differential characteristics of the original signal. Finally, the quantum enhanced features are fused with the original temporal features and input into a lightweight convolutional neural network for classification, outputting the probability distribution of signal quality levels.

[0034] This training method utilizes mathematical geometric features to characterize the dynamic nature of signals and enhances feature separability through quantum transformations constrained by physical laws. Ultimately, it achieves accurate and robust detection of motion artifacts in blood pressure signals across various scenarios, from static to strenuous exercise. It significantly improves signal quality detection accuracy and generalization ability under motion interference. The algorithm's lightweight design makes it suitable for embedded deployment, providing crucial technical support for reliable, continuous, and non-invasive blood pressure monitoring in wearable devices.

[0035] Furthermore, such as Figure 2 As shown, the preprocessing and window segmentation include the following steps: Step S11: The original PPG signal and the original motion reference signal are synchronously sampled using the same master clock, and the sampling frequency is set to be no less than 100 Hz. Step S12: Perform a first bandpass filter on the original PPG signal, with a passband frequency range of 0.5 Hz to 8.0 Hz; Step S13: Perform a second bandpass filter on the original motion reference signal, with a passband frequency range of 0.1 Hz to 20.0 Hz; Step S14: Perform window segmentation on the filtered PPG signal and the motion reference signal to obtain the original multi-channel signal segments; Step S15: Normalize the original multi-channel signal segment to eliminate signal amplitude scaling and DC offset caused by individual and equipment differences, and obtain the multi-channel signal segment.

[0036] Specifically, in step S11, the original PPG signal and the original motion reference signal are synchronously acquired and digitized. The specific operation process is as follows: Step S11.1, Definition of Mathematical Symbols Continuous time.

[0037] Sampling period (unit: seconds) The sampling frequency.

[0038] Discrete-time index , .

[0039] At discrete time points n The acquired raw PPG signal amplitude (usually a direct reading from the analog-to-digital converter ADC). 、 and At discrete time points n The amplitude of the acquired triaxial raw acceleration signal (ACC signal).

[0040] Step S11.2, Hardware Synchronization Mechanism To ensure that the PPG signal and the motion reference signal are strictly aligned in time, the following synchronization strategy is adopted: the PPG sensor and the inertial measurement unit (IMU, including accelerometer) share the same master clock for sampling, or a precise timestamp under a unified reference is added to the two sets of data.

[0041] Through the above synchronization mechanism, it is fundamentally guaranteed that , 、 and These are physical quantity measurements taken at exactly the same moment. This is a prerequisite for subsequent calculations of key characteristics such as the time delay and cross-correlation coefficient between PPG and ACC.

[0042] Step S11.3, Sampling Parameter Setting Based on the Nyquist theorem and physiological signal characteristics, the main energy of the PPG signal is concentrated below 20Hz. To retain sufficient harmonic information while also considering subsequent processing requirements, the sampling frequency of the PPG signal is set accordingly. Simultaneously, the ACC signal needs to capture human motion (typically not exceeding 20Hz), and the sampling frequency of the ACC signal is set for analyzing high-frequency impact components. To simplify system design and data alignment, it is preferable to set both to the same sampling frequency, such as 100 Hz or 125 Hz.

[0043] In steps S12 and S13, the original PPG signal and the original motion reference signal are filtered. The purpose of filtering is to retain physiological information related to heart rate and blood pressure while removing baseline drift, power line interference, and high-frequency noise as much as possible. The specific filtering process is as follows: Step S12.1, Definition of Mathematical Symbols : No. k The raw PPG signal for each window.

[0044] : No. k The raw ACC signal for each window (processed independently for each axis).

[0045] 、 The unit impulse responses of the digital filters designed for PPG and ACC, respectively. m For filter tap index, This represents a linear convolution operation.

[0046] Step S12.2, PPG signal filter A fourth-order Butterworth IIR bandpass filter was used to filter the PPG signal segment, with a passband cutoff frequency ranging from 0.5 Hz to 8.0 Hz. The lower cutoff frequency of 0.5 Hz was used to remove baseline drift caused by changes in respiration and body temperature (typically below 0.5 Hz); the upper cutoff frequency of 8.0 Hz was used to preserve heart rate (typically 0.8–3.0 Hz) and its main harmonics, while removing high-frequency electromyographic noise and ambient light interference.

[0047] Step S12.3, ACC signal filter A fourth-order Butterworth IIR bandpass filter was used to filter the ACC signal segments of each axis, with a passband cutoff frequency ranging from 0.1 Hz to 20.0 Hz. The lower cutoff frequency of 0.1 Hz was used to remove the gravity component changes caused by slow changes in the equipment's posture; the upper cutoff frequency of 20.0 Hz was used to retain the main limb motion frequency components while suppressing high-frequency vibration noise.

[0048] Step S12.4 Filtering operation The above filtering operation can be expressed as: (1), (2), Among them, the triaxial acceleration signal 、 、 Perform the same filtering operation on each.

[0049] In practical digital signal processing, zero-phase filtering (e.g., the FLTFFIT function) is preferred to avoid phase shift introduced by the filter, thereby ensuring that waveform characteristics are not distorted on the time axis. This is of great significance for subsequent calculations of characteristics such as pulse wave propagation time.

[0050] Step S14 is used to implement signal segmentation and windowing. In order to convert an infinitely long continuous signal stream into finite-length segments that can be processed one by one, this embodiment performs window segmentation on the synchronously acquired signal. This is the basic operation for realizing real-time processing.

[0051] Step S14.1, Definition of Mathematical Symbols Define the following window-related parameters: L Window length, in seconds. Based on the quasi-periodicity of the blood pressure waveform and the continuity of the exercise pattern, this embodiment selects... L =8 seconds. This window length is sufficient to include a large number (approximately 8 to 12) of cardiac cycles while ensuring real-time processing (with manageable latency).

[0052] Window length, in units of sample points. If L = 8 seconds, ， but point.

[0053] R Window overlap ratio, usually set to R =0.5. This ensures that no signal feature is truncated or lost because it falls exactly at the edge of the window, and increases the continuity of the data.

[0054] H : Slide step size points .like point, R =0.5, then H =400 points.

[0055] k Window index, .

[0056] : No. The first window One PPG signal point. .

[0057] Step S14.2, Window Splitting Operation For any signal (represent or any axis ） , its first k The data for each window is: for (3), For example, the 0th window contains S [0] to S

[799] , the first window contains S

[400] to S

[1199] , and so on. This operation generates two aligned sequence of signal windows: and .

[0058] In step S15, the signal is normalized. Normalization aims to eliminate signal amplitude scaling and DC offset caused by individual differences (such as skin color and subcutaneous fat thickness) and device differences (such as LED intensity and detector gain), so that the model focuses on the shape and variation of the signal rather than its absolute magnitude.

[0059] Step S15.1, Definition of Mathematical Symbols : No. The PPG signal after window filtering (length is...) (vector).

[0060] :In this window The arithmetic mean.

[0061] :In this window The standard deviation.

[0062] Step S15.2, details ( standardization): For the first Each signal point within a window undergoes the following transformation: (4), (5), in, It is a very small constant (e.g.) This is used to prevent division by zero errors when the standard deviation is zero. Similarly, the accelerometer signal undergoes the same normalization process.

[0063] After the above normalization operation, the signal sequence within each window is converted into a standardized sequence with a mean of 0 and a standard deviation of 1. This processing has the following technical effects: Amplitude scale normalization: ensures that signals collected from different windows, different users, and different devices are comparable in amplitude scale; Accelerated algorithm convergence: It eliminates the huge differences in numerical scale of features in various dimensions, which greatly accelerates the convergence speed of subsequent gradient descent optimization algorithms; Improve generalization ability: make the model focus more on the morphological characteristics and variation patterns of the signal, rather than the absolute amplitude, thereby improving the model's generalization ability across different individuals and devices.

[0064] After the above five interconnected sub-steps, for each window index... k This yields a clean, standardized set of input pairs, which are then passed to step S20 for feature extraction. (6), in, Furthermore, all signals have been synchronized, filtered, and normalized. The normalized PPG data and the ACC signal groups for each axis serve as the common starting point for subsequent calculus transformations, quantum coding, and deep learning classification.

[0065] In one embodiment of the present invention, the construction of the static feature vector and temporal feature matrix of the multi-channel signal segment is as follows: Figure 3 As shown, it includes the following steps: Step S21: Calculate the first-order difference and the second-order difference for the PPG signal and the motion reference signal in the multi-channel signal segment respectively to obtain the differential sequence, and extract the first set of scalar features based on the differential sequence. The first set of scalar features includes at least one of standard deviation, kurtosis and zero crossing rate. Step S22: Calculate the absolute integral and square integral within the sliding window for the PPG signal and motion reference signal in the multi-channel signal segment respectively to obtain the integral sequence, and extract the second set of scalar features based on the integral sequence. The second set of scalar features includes at least one of mean, coefficient of variation and root mean square. Step S23: Calculate the normalized cross-correlation function between the differential sequence of the PPG signal and the differential sequence of the motion reference signal, extract the maximum value of the cross-correlation coefficient and its corresponding time delay, and calculate the energy ratio of the PPG signal and the motion reference signal, which are used as the third set of scalar features. Step S24: Concatenate the first group of scalar features, the second group of scalar features, and the third group of scalar features into a static feature vector, and retain one or more of the original PPG signal, the original motion reference signal, and its first-order differential sequence, and arrange them in time sequence into a time-series feature matrix.

[0066] Specifically, within the signal window generated in step S10, which has undergone synchronization, filtering, and normalization, a set of calculus features is computed in parallel. This set of features can characterize the dynamic properties of the signal from multiple scales and dimensions, forming a physical semantic bridge connecting the original signal and advanced intelligent algorithms, aiming to sensitively distinguish between physiological pulse fluctuations and motion artifacts.

[0067] Input: number kStandardized signals for each window (output from step S10): (7), Motion artifacts in signals primarily manifest as non-physiological high-frequency jitter, abrupt changes, and low-frequency baseline drift. To address these characteristics, differential operations amplify instantaneous signal changes, exhibiting high sensitivity to abrupt changes and jitter; integral operations reflect the cumulative trend of the signal, demonstrating high sensitivity to baseline drift and sustained motion. Combining these two methods allows for the quantification of signal characteristics along two orthogonal dimensions: the rate of change of signal and energy accumulation, thereby effectively distinguishing physiological pulse fluctuations from motion artifacts.

[0068] Differential feature set calculation: In step S21, this step estimates the rate of change of the signal by calculating the first and second differences to enhance the sensitivity to rapid, non-physiological abrupt changes introduced by motion artifacts.

[0069] Mathematical symbols and calculation: First-order forward difference (approximate first-order derivative): (8), (9), in, Similarly, using discrete-time indexing, the first-order difference of the accelerometer's signals along each axis can be calculated. , 。 Physical meaning: The magnitude directly reflects the PPG waveform at time [time]. n The instantaneous slope of the pulse wave is affected by artifacts from intense motion, which can cause the value to spike far beyond the normal pulse wave waveform.

[0070] Second-order difference (approximates the second derivative, capturing curvature changes): (10) Physical meaning: This reflects the acceleration of signal changes. The second derivative of a physiological pulse waveform changes relatively smoothly, while impact motion artifacts cause it to oscillate violently. The following statistical features are extracted from the above difference sequences to constitute the static features describing the overall dynamics of the window: Standard deviation (Std) is used to describe the magnitude of change. (11), in, differential sequence The arithmetic mean, Std represents the sequence length. The larger the Std value, the more severe the signal jitter within the window.

[0071] Kurtosis: Describes the sharpness of a distribution.

[0072] (12), in, denoted by μ, the expected probability. A high kurtosis value indicates that the distribution of the difference sequence has a "heavy-tailed" characteristic, meaning that there are many extreme values ​​far from the mean (corresponding to sudden and severe disturbances). Here, μ is the mean.

[0073] Zero Crossing Rate (ZCR): (13) Where N is the sequence length, As an indicator function, a high zero-crossing rate indicates frequent and rapid fluctuations in the signal, a typical characteristic of high-frequency noise. Output: For PPG... D and DD Sequences, and three-axis ACC D For each sequence, calculate the three statistics mentioned above. This will generate: (1+1+3) signals × 3 features = 15 differential statistical features.

[0074] In step S22, the integral feature family calculation is performed (capturing cumulative trends and energy). The integral operation is approximately a summation, reflecting the cumulative effect or energy of the signal over a time period, and is sensitive to continuous motion interference and baseline drift.

[0075] Mathematical symbols and calculation: Sliding window absolute integral (representing the absolute change in the signal): Define child window length M (For example, corresponding to 2 seconds:) ); (14) According to Passevar's theorem, this directly corresponds to the signal energy in the time domain, leading to feature extraction: Mean and coefficient of variation: Calculation The mean of the sequence ( The ratio of the standard deviation to the mean (coefficient of variation) and the ratio of the standard deviation to the mean. The mean reflects the overall activity level, while the coefficient of variation reflects the degree of fluctuation in activity level. Both typically increase significantly during exercise.

[0076] Root mean square (RMS): The square root of the integral of the square of the entire window, a classic representation of energy. (15) Output: Calculate the PPG and triaxial ACC values ​​respectively. sequential This will generate: 4 signals × 3 features = 12 integral statistical features.

[0077] In step S23, cross-calculus feature calculation is performed (to capture intermodal coupling relationships). This is key to distinguishing physiological blood flow changes from external motion interference. Under physiological conditions, PPG and ACC changes are weakly correlated; however, when motion artifacts dominate, their changes are highly correlated.

[0078] Mathematical symbols and calculation: Cross-correlation coefficients and time delay: Calculate the PPG differential signal Differential signals of each axis ACC Normalized cross-correlation coefficient (NCC) sequence: (16) Take the maximum absolute value of the sequence and its corresponding time delay. τ : (17) (18) high This indicates that the mutations in PPG and ACC are highly synchronized, providing strong evidence for motion artifacts. It can help determine the propagation path of artifacts.

[0079] Energy ratio: (19) This ratio decreases significantly during strenuous exercise (when ACC energy surges).

[0080] Output: Calculate PPG and the three ACC axes separately. and , plus a total This generates 3 axes. 2 features + 1 = 7 cross features.

[0081] In step S24, all the above scalar features are concatenated into a comprehensive static feature vector, while the key original time-series signals are retained for subsequent CNN processing.

[0082] static feature vector : (20) This is a (15+12+7)=34 dimensional vector. It highly summarizes the dynamic characteristics, energy distribution, and multimodal coupling relationships of the signal within the entire window.

[0083] Temporal feature matrix F t (CNN used in step S40): (twenty one), This is a size of The matrix (assuming four key timing channels are selected: PPG, PPG difference, one-axis ACC, and the ACC difference) preserves the complete original waveform and its first-order differential form, which is used by CNN to mine local spatiotemporal patterns.

[0084] For each window k, output a tuple: .

[0085] A 34-dimensional static feature vector will be used as the input for step 3 (quantum feature enhancement).

[0086] : L points x4 The temporal feature matrix will be used as one of the inputs for step S40 (CNN classification).

[0087] All calculations in step S20 strictly depend on the normalized, aligned signal window provided in step S10. and The calculated differential characteristic Directly from Difference and integral characteristics It is a local summation of their absolute values, and the cross feature NCC is the correlation operation between their differential sequences. This is a concise summary of the signal dynamic characteristics in step S10. This is the ordered arrangement of the signal and its direct derivative from step S10. This step transforms the original signal into statistical features that contain clearly defined physical meaning. （ ） It also preserves the original waveform information. （ ） The two representations pave the way for subsequent mixed processing procedures.

[0088] Furthermore, the quantum feature extraction module includes a classical encoder, an angle encoding layer, a variable quantum circuit, a measurement layer, and a linear layer.

[0089] In another embodiment of the present invention, the static feature vector is input into the quantum feature extraction module for quantum feature extraction processing to obtain a quantum enhanced feature vector, such as... Figure 4 As shown, it includes: Step S31: Map the static feature vector to rotation angle parameters using the classic encoder; Step S32: The rotation angle parameter is loaded onto the rotation gate of the qubit through the angle encoding layer to generate the input quantum state of the quantum circuit; Step S33: Transform the input quantum state using the parametric quantum gates and entanglement gates in the variable quantum circuit to obtain the final quantum state; Step S34: Measure the final quantum state through the measurement layer, and perform a linear transformation on the measurement result through the linear layer to obtain the quantum enhancement feature vector.

[0090] Specifically, the core objective of step S30 is to transform the static feature vector generated in step S20. F s Mapping to a high-dimensional quantum Hilbert space, the parallelism and superposition properties of quantum computing are utilized to enhance the separability of features, thereby better distinguishing signal quality. Simultaneously, differential constraints are introduced to ensure that changes in quantum features remain consistent with the physical dynamics (differential features) of the original signal, improving the physical interpretability and generalization ability of the features.

[0091] Input: The static feature vector output from step S20 (Dimension c), and the differential features from step S21 (used for constraints, such as...) Output: Quantum-enhanced eigenvectors (dimension) m ,For example ).

[0092] Classical quantum interface encoding (encoding classical eigenvectors (static eigenvectors) into quantum states) is necessary because quantum computing units process qubits, requiring the classical eigenvectors to be encoded into quantum states. The initial conditions are encoded into quantum states. Here, a combination of trainable classical neural networks and angle coding is used to achieve adaptive feature compression and encoding.

[0093] Mathematical symbols and operations: Suppose that a quantum system has n qubits (e.g.) n =8). The ground state of each qubit is and .

[0094] Trainable Classic Encoder: First, a shallow, fully connected neural network is used to train the 34-dimensional classic feature vector. F s Compress and map to n 3D rotation angle vector θ .

[0095] Network structure: Input layer (34) > Fully connected layer (16) > ReLU activation > Fully connected layer ( n > Sigmoid activation > Output multiplied by π 0.

[0096] Formula expression: (twenty two), (twenty three), in, ， This is the weight matrix. ， For bias vectors, The sigmoid() function compresses the output to... Multiply by π The angle vector is then obtained. .

[0097] First, angle encoding: This involves encoding the angle vector... θ Each component is encoded into the quantum state of the corresponding qubit. For the ... Apply a rotation gate around the Y-axis to each qubit. : (twenty four), entire n The initial quantum state of a qubit system is the direct product of the states of each qubit (separable state): (25), Objective: A trainable encoder can learn how to optimally compress and map classical features into quantum space, providing an information-rich initial state for subsequent quantum transformations. Angular coding is a simple and efficient coding method that fully utilizes the continuity of quantum rotating gates.

[0098] Next, the variable quantum circuit (VQC) transformation: Variable quantum circuits consist of a series of parameterized quantum gates whose parameters can be adjusted using a classical optimizer to achieve the desired quantum state transformations. A layered hardware-efficient implementation is designed.

[0099] Circuit structure: The circuit consists of L Layers (e.g.) L =4) Repetitive structure. Each layer contains two sub-layers: Single-qubit rotation sublayer: Applying a rotational layer to each qubit in sequence y shaft and z The rotation of the axis can be trained.

[0100] For the I layer( ), for the first qubits, with gate sequence applied: .

[0101] The operation of this sub-layer is: (26) Entangled sublayer: Employs linear nearest-neighbor coupling, i.e., coupling adjacent qubits Apply a CNOT gate to create entanglement.

[0102] The operation of this sub-layer is as follows: (27) Therefore, the overall transformation of the first layer is: (28) The transformation of the entire variable quantum circuit is as follows: (29) in It is the set of all trainable parameters.

[0103] initial state After passing through the variable quantum circuit, the evolution yields the final state: (30) The final state is n The entangled state of a quantum bit carries characteristic information that has undergone complex nonlinear transformations.

[0104] The goal is that, through parameterized quantum gates and entanglement operations, quantum circuits can perform nonlinear transformations on encoded classical information in an exponentially high-dimensional Hilbert space, potentially finding feature representations that make it easier to distinguish different categories (good / medium / poor signal quality), so as to achieve at least a first quality level, a second quality level, and a third quality level (i.e., good, medium, and poor quality levels).

[0105] Furthermore, quantum measurement and feature readout: Measurement scheme: Select a set of Pauli operators' expected values ​​as features. For each qubit, measure the expected value of its Pauli Z operator: (31), in Z i Pauli Z Matrix action in the first i One qubit is represented by an identity matrix, and the expected value is between [-1, 1]. It is a quantum state.

[0106] Simultaneously, in order to capture the correlation between qubits, the Pauli score of each pair of qubits is measured. ZZ Operator expectation: (32), Feature vector construction: Concatenate all single-bit and two-bit expected values ​​into a single primitive quantum eigenvector: (33), Its dimensions are n+C(n,2) .when n When the value is 8, the dimension is 8 + 28 = 36.

[0107] To further compress and integrate information, Mapped through a trainable fully connected layer (linear layer) to m Dimensions (e.g.) m =16): (34), in ， These are trainable parameters.

[0108] Objective: The Pauli expectation measurement provides statistical information about quantum states in various directions, including single-qubit and two-qubit correlations, thus forming a rich feature representation. Subsequent linear layers further fuse these features and reduce dimensionality to suit downstream tasks.

[0109] Furthermore, perform quantum training with differential constraints: When training the parameters of a variable quantum circuit, in addition to the classification loss, a consistency constraint loss based on differentiation is introduced, so that the variation pattern of quantum features is associated with the differential features of the original signal, thereby injecting physical priors into the quantum model.

[0110] Constraint Design: In step S20, the statistical characteristics of the first-order difference sequence of the PPG signal are calculated, such as its standard deviation. This reflects the intensity of fluctuations in the original signal. It is hoped that the intensity of changes in the quantum characteristics will remain consistent with this statistic.

[0111] Let the current window index be k Its quantum enhancement feature is (16-dimensional). Considering the characteristic changes of a continuous window, the square of the Euclidean norm of its difference (approximate time derivative) is calculated as a measure of the intensity of the quantum characteristic change: (35), hope Differential features from step S20 ，k They are positively correlated. However, due to the different scales, their standardized values ​​are constrained to be as close as possible.

[0112] Differential constraint loss function: For a batch of B samples, calculate the standardized mean squared error: (36) in, ， It is in this batch The mean and standard deviation, ， It is in this batch The mean and standard deviation.

[0113] Overall loss function and training: Suppose that the classification prediction obtained after passing through the classification network in the subsequent step S40 is: p The real label is y (Signal quality level: Good, Medium, Poor). The classification loss uses cross-entropy loss. (37) The total loss is: (38), in The overshoot parameter (e.g., set to 0.1).

[0114] The training process is as follows: Forward propagation: static feature vector Classic encoder > Angle encoding > Variable quantum circuit > Measurement > Linear layer acquisition Enhance the eigenvector; With step S20 After fusion, the result is input into the CNN (Convolutional Neural Network) in step S40 to obtain the prediction. ; Calculate loss ; All trainable parameters, including classical encoder parameters and variable quantum circuit parameters, are updated using gradient descent. Linear layer parameters ， And CNN parameters.

[0115] Differential constraint loss forces the intensity of changes in quantum features to remain consistent with the differential statistics of the original signal, essentially injecting prior knowledge from classical signal processing (motion artifacts increase the differential features of the signal) into the quantum machine learning process. This makes the quantum feature enhancement process more in line with physical laws, contributing to improved model generalization ability and interpretability.

[0116] In summary, for each window k The trained quantum feature enhancement module outputs a quantum enhanced feature vector. (16-dimensional). This vector not only contains a quantum high-dimensional mapping of classical features, but its dynamic changes are also related to the differential properties of the original signal, thus making it more likely to contain discriminative information related to motion artifacts.

[0117] In one embodiment of the present invention, the feature fusion of the quantum-enhanced feature vector and the temporal feature matrix is ​​performed, such as... Figure 5 As shown, it includes: Step S41: Copy the quantum enhancement feature vector along the time dimension to generate a quantum feature matrix with the same time length as the temporal feature matrix and with each row being identical; Step S42: Concatenate the temporal feature matrix and the quantum feature matrix along the feature channel dimension to obtain a fused feature matrix; Step S43: Convert the fused feature matrix into a three-dimensional tensor.

[0118] In other words, the goal of this step is to transform the time-series feature matrix from step S20. F t and the quantum-enhanced eigenvector of step S30 F q Effective feature fusion is performed, and joint spatiotemporal features are extracted using a lightweight convolutional neural network (CNN), ultimately outputting the probability distribution of signal quality categories. This step is the core of the classification decision, and its design must balance the rationality of feature fusion with the computational efficiency on embedded devices.

[0119] Input: The time-series feature matrix from step S20 ,in (In the example) , First-order difference, , First-order difference (four channels) (8-second window, 100 Hz sampling).

[0120] Quantum-enhanced eigenvectors from step S30 ,in .

[0121] Output: Probability distribution of signal quality categories These correspond to three quality levels: good, medium, and poor, and satisfy the following conditions: .

[0122] Specifically, the key to this step is the effective fusion of static quantum enhancement features with dynamic temporal features. This invention employs a "channel replication and splicing" strategy, using global quantum features as contextual information for local time points.

[0123] The mathematical symbols and operations are as follows: Quantum feature serialization: quantum eigenvectors (dimension) m Replicating along the time dimension to construct a [structure / structure] with [other components]. Matrix with the same time length : (39) That is, matrix Every line is the same Transpose. This means that the entire time window shares the same quantum enhancement feature, providing global, quantum-transformed contextual information for each time point.

[0124] Channel dimension splicing: Time series feature matrix and quantum characteristic matrix The features are concatenated in the second dimension (feature channel dimension) to obtain the fused feature matrix. : (40) Where || indicates splicing along the column direction. ， hour, X The channel dimension is 20.

[0125] Tensor Reshaping: To adapt to the input format of a one-dimensional CNN, the matrix X Convert to 3D tensor ,in B Batch size (during training) B >1, during reasoning B =1). In neural network frameworks, the channel dimension is usually placed before the length dimension, i.e., the shape is... ,in For the number of channels, The sequence length is given.

[0126] Physical meaning: This fusion method enables the CNN to perform at every time point.n Above, not only can the original time-series signal and its derivative (from...) be observed, but also... Furthermore, quantum enhancement features representing the global properties of the entire window can be observed (from...). This helps the network determine whether local fluctuations are part of a global motion artifact or local physiological changes.

[0127] Lightweight 1D Convolutional Neural Network Architecture Design: Design a lightweight network with depthwise separable convolutions, dilated convolutions, and global pooling to efficiently extract features and perform classification with limited computational resources.

[0128] Define the input tensor as ,in .

[0129] Layer 1: Depthwise Separable Convolution Objective: To significantly reduce the number of parameters and computational load, while separating feature learning within and across channels.

[0130] The operation is as follows: Depthwise Convolution: Each input channel undergoes independent one-dimensional convolution. (Using...) The convolution kernel, with the number of groups set to the number of input channels. C No bias. The number of output channels remains the same. C .

[0131] (41), in These are the weights of the depthwise convolution. Output. .

[0132] Pointwise convolution: using Convolution (i.e., fully connected across channels) increases the number of channels from Expand to .

[0133] (42), in Output ,in .

[0134] Follow-up actions: H Perform batch normalization and ReLU activation.

[0135] Layer 2: Dilated Convolution Objective: To expand the receptive field and capture longer-term temporal dependencies without increasing parameters or reducing resolution.

[0136] Operation: Use a standard one-dimensional convolution, but with an expanded kernel. Let the kernel size be... Expansion rate Number of output channels Step size = 1, no padding.

[0137] (43), in Due to the expansion, the actual interval between input locations involved in the calculation is... d Output length .

[0138] Subsequent steps: Batch normalization. ReLU activation.

[0139] Layer 3: Global Average Pooling Objective: To aggregate the entire time dimension of each channel into a scalar, significantly reduce the number of parameters, provide channel-level global information, and maintain a certain degree of translation invariance.

[0140] Operation: Take the average value for all time points for each channel.

[0141] (44), Output That is, each sample gets a 3D eigenvectors.

[0142] Layers 4 and 5: Fully Connected Classifier Objective: To map high-level features to class probabilities.

[0143] The operation is as follows: Fully connected layer 1: Maps 32-dimensional features to 16 dimensions, followed by... ReLU Activation and Dropout (Discard rate) p=0.5 To prevent overfitting: (45) (46).

[0144] Fully connected layer 2 (output layer): maps 16 dimensions to 3 dimensions (corresponding to three categories), followed by a Softmax activation function. (47) (48) The final output is the probability that each sample belongs to one of the three quality levels. .

[0145] Training, optimization, and inference: The entire model (the quantum feature enhancement module in step S30 and the CNN module in step S40) is jointly trained end-to-end. The overall loss function is: (49) in, It is the classification cross-entropy loss, for batch data: (50), This is the true label (one-hot encoded) of sample b. It is the differential constraint loss. It is a hyperparameter (e.g., set to 0.1).

[0146] Optimization: Gradient descent is performed using the Adam optimizer to update all trainable parameters, including: the classical encoder parameters from step S30 and the variable quantum circuit parameters. Quantum characteristic linear layer parameters And the CNN parameters in step 4 (convolutional layer weights, fully connected layer weights, etc.).

[0147] The gradient of a quantum circuit can be estimated on a classical simulator by automatic differentiation or by using methods such as parameter shifting rules on real quantum hardware.

[0148] The learning rate can be set to 0.001, and a learning rate decay strategy may be used.

[0149] reasoning: After training, during inference on the deployment (e.g., a wearable device), for each new 8-second window of data, the forward propagation steps S10, S20, S30, and S40 are executed sequentially to obtain the probability vector. The category with the highest probability is selected as the quality level of the prediction. (51), At the same time, probability value The probability of the "good" category itself (especially the probability of the "good" category) can be used as a confidence score for subsequent decision-making (step S50).

[0150] Output: For each processing window This step outputs a three-dimensional probability vector. , indicating the confidence level of the window signal quality as good, medium, or poor, respectively.

[0151] Connection with preceding and following steps: Input dependencies: Input for this step and It comes directly from steps S20 and S30, and its calculation process is closely linked.

[0152] Output propagation: the probability vector of the output It will be directly passed to the decision module in step S50 for final device feedback and user interaction.

[0153] At this point, the intelligent processing flow from raw data to quality classification of the signal has been completed. Step S50 will make a real-time decision based on this classification result and in combination with the specific application scenario.

[0154] According to one embodiment of the present invention, the convolutional neural network is composed of a depth-separable convolutional layer, a dilated convolutional layer, a global average pooling layer, and at least one fully connected layer connected in series.

[0155] Preferably, the gradient descent method uses the Adam optimizer to perform parameter iteration; for the trainable parameters of the variable quantum circuit, the gradient is calculated using parameter shift rules or automatic differentiation.

[0156] Furthermore, gradient descent is performed using the Adam optimizer to update all trainable parameters, including: the classical encoder parameters from step S30 and the variable quantum circuit parameters. Quantum characteristic linear layer parameters ， The learning rate can be set to 0.001, and a learning rate decay strategy can be used. This includes the CNN parameters (convolutional layer weights, fully connected layer weights, etc.) from step S40.

[0157] The gradient of a quantum circuit can be estimated on a classical simulator by automatic differentiation or by using methods such as parameter shifting rules on real quantum hardware.

[0158] Step S50 is the final output of the signal quality assessment process. Its goal is to transform the signal quality probability distribution output in step S40 into specific, actionable system actions, and to provide necessary and restrained feedback guidance to the user. The core design principles are to ensure data reliability, maintain user experience, and minimize interference.

[0159] In a second aspect, the present invention provides a variable quantum circuit training device 100, such as... Figure 6 As shown, it includes: The data acquisition module 110 is used to acquire a training dataset, which contains multiple multi-channel signal segments. Each multi-channel signal segment is obtained by preprocessing and window segmentation of the synchronously acquired raw PPG signal and raw motion reference signal, and each multi-channel signal segment corresponds to a real label. The feature construction module 120 is used to construct the static feature vector and temporal feature matrix for each of the multi-channel signal segments; The quantum feature extraction module 130 is used to perform quantum feature extraction processing on the static feature vector to obtain a quantum enhanced feature vector; The fusion classification module 140 is used to fuse the quantum-enhanced feature vector with the temporal feature matrix, and input the fused features into a convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. The loss calculation module 150 is used to calculate the differential consistency constraint loss based on the change intensity of the quantum enhancement feature vector corresponding to adjacent time windows and the standard deviation of the original PPG signal difference sequence within the corresponding window, calculate the classification loss based on the probability prediction result and the true label, and perform a weighted summation of the classification loss and the differential consistency constraint loss to obtain the total loss. The parameter update module 160 is used to iteratively update the trainable parameters of the quantum feature extraction module and the convolutional neural network classifier using the gradient descent method with the goal of minimizing the overall loss, thereby completing the training.

[0160] The variable quantum circuit training device provided by the present invention can implement all the processes of the above-described variable quantum circuit training method embodiments and achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0161] According to embodiments of the present invention, a computer program and a computer program product are also provided. When the computer program is executed by a processor, it can implement the variable quantum circuit training method described in any of the foregoing embodiments of the present invention and achieve the same technical effects, which will not be repeated here.

[0162] This invention also provides a computer-readable storage medium, which can be implemented using one or more computer-readable media in combination. The computer-readable storage medium is a tangible storage medium, including but not limited to: electrical connection components having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, optical disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, and any combination of the above devices. This computer-readable storage medium is used to store computer programs, which can be invoked and run by an instruction execution system, device, or equipment.

[0163] The program code in the storage medium can be transmitted via any suitable method, such as wireless, wired, optical fiber, or radio frequency.

[0164] The computer program code used to implement the technical solution of this invention can be written individually or in combination using object-oriented programming languages ​​such as Java, Smalltalk, and C++, or conventional procedural programming languages ​​such as C. The program code can run entirely on a local device, partially on a local device, run independently as a software package, or partially / fully on a remote computer or terminal. When a remote computer is involved, the remote computer can be accessed via a local area network (LAN), a wide area network (WAN), or via an internet service provider to achieve communication. In a fifth aspect, this invention proposes a method for detecting motion artifacts in blood pressure signals, based on a quantum feature extraction module and a convolutional neural network classifier trained using the training method described in any of the above embodiments, such as... Figure 7 As shown, it includes the following steps: Step S100: Acquire the synchronized raw PPG signal and raw motion reference signal, perform preprocessing and window segmentation to obtain the multi-channel signal segment to be detected; Step S200: Construct the static feature vector and temporal feature matrix of the multi-channel signal segment to be detected; Step S300: Input the static feature vector into the trained quantum feature extraction module to obtain the quantum enhanced feature vector; Step S400: Perform feature fusion between the quantum-enhanced feature vector and the temporal feature matrix, and input the fused features into the trained convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window; Step S500: Perform a hierarchical decision based on the probability prediction result. The hierarchical decision includes at least adopting the PPG signal of the current window for blood pressure calculation, adopting it after marking it for blood pressure calculation, or rejecting it for blood pressure calculation, and triggering user feedback when rejecting it.

[0165] It should be noted that in step S500, the input is: the current window of step S400. k Signal quality probability vector: ,in and These represent the probabilities of different quality levels, and satisfy the following conditions: .

[0166] Output: Final decision ; User feedback instructions 。

[0167] (I) Deterministic Decision Making and Confidence Mapping By following clearly defined rules, a deterministic decision action is generated. Decision rules aim to transform the uncertainty of probability into clear and reliable operational instructions.

[0168] According to one embodiment of the present invention, the probability prediction result includes a first quality level (good), a second quality level (medium), and a third quality level (poor), such as... Figure 8 As shown, the step of performing hierarchical decision-making based on the probability prediction results includes: Step S410: Set a high confidence threshold and a low confidence threshold, wherein the high confidence threshold is greater than the low confidence threshold; Step S420: Take the category corresponding to the maximum value among the probabilities of the first quality level, the second quality level, and the third quality level as the predicted main category; Step S430: When the predicted main category is the first quality level and its probability is greater than or equal to the high confidence threshold, the decision is to adopt the current signal; Step S440: When the predicted main category is the first quality level and its probability is between the low confidence threshold and the high confidence threshold, or when the predicted main category is the second quality level, or when the predicted main category is the third quality level and its probability is less than the low confidence threshold, the decision is to mark and then adopt. Step S450: When the predicted main category is the third quality level and its probability is greater than or equal to the high confidence threshold, the decision is to reject the current signal.

[0169] Furthermore, the decision-making rule is defined as follows: Set two fixed thresholds: high confidence threshold (e.g., 0.7) and low confidence threshold (e.g., 0.3), satisfying .

[0170] Decision function Defined as: (52), The semantics of the decision-making action and the subsequent system-level processing are as follows: Accept: Semantics: The PPG signal quality of the current window is determined to be highly reliable and can be used for accurate blood pressure calculation.

[0171] System Action: The raw PPG signal (or a preprocessed but unnormalized signal) from this window is transmitted to the device's built-in blood pressure calculation algorithm. The calculated systolic blood pressure, diastolic blood pressure, heart rate, and other values ​​will be marked as "high confidence" measurement results for display on the device interface, local storage, and cloud synchronization.

[0172] Accept With Warning: Semantics: The quality of the PPG signal in the current window is uncertain and may be affected by slight motion interference, which may lead to errors in the blood pressure calculation results.

[0173] System Action: The PPG signal from this window is also transmitted to the core blood pressure calculation algorithm. However, the calculation result will be appended with a "low confidence" label. At the application layer, this type of data will: When displayed on the device screen, it may be accompanied by an exclamation mark icon or a change in the color of the numerical value.

[0174] It is not included in the real-time display of "current blood pressure" readings (if the device has this function), but can be recorded in historical trends.

[0175] When calculating average blood pressure, it can be given a lower weight or excluded.

[0176] Reject: Semantic: The quality of the PPG signal in the current window is unreliable, and any blood pressure calculation based on it is very likely to produce misleading results.

[0177] System action: The PPG signal in the current window is discarded without triggering any blood pressure calculation process. The system will then activate the user feedback mechanism to guide the user to improve measurement conditions.

[0178] Basis for threshold selection: threshold and It is determined during the model development phase by balancing the true positive rate (correctly identifying high-quality signals) and the false positive rate (incorrectly accepting low-quality signals) on an independent validation set. Its goal is to maximize system usability within a clinically acceptable error range.

[0179] According to one embodiment of the present invention, when the decision is to adopt the current signal, the PPG signal of the current window is transmitted to the blood pressure calculation algorithm, and the calculated blood pressure value is marked as a high-confidence measurement result.

[0180] According to another embodiment of the present invention, when the decision is to adopt the marked value, the PPG signal of the current window is transmitted to the blood pressure calculation algorithm, and the calculated blood pressure value is marked as a low confidence measurement result.

[0181] According to another embodiment of the present invention, when the decision is to reject the current signal, the PPG signal of the current window is discarded, the blood pressure calculation process is not triggered, and the user feedback mechanism is started.

[0182] Preferably, the triggering of user feedback includes: Obtain the user's current state, which includes both active and inactive states; When the decision is to reject the current signal and the user is active, a vibration alert is triggered. A visual cue is triggered when the decision is to reject the current signal and the user is stationary.

[0183] Furthermore, context-aware minimal interference feedback.

[0184] The goal of user feedback is to provide prompts only when necessary (signals are rejected) and effective (allowing the user to take action), in a manner that minimizes user disruption. This design employs feedback logic based on decision outcomes and a simple state machine.

[0185] Feedback rule definition: Define feedback function ,in It is a simple current user state that the system determines in real time based on accelerometer signals (for example, Active means there has been significant movement in the past few seconds, Quiet means relatively still).

[0186] (53), The physical execution of the feedback action includes: after triggering a vibration or visual cue, the system enters a preset silence period during which user feedback is prohibited from being triggered again. Specifically: Vibrate alert trigger scenario: The signal is rejected and the user is active. This usually means that motion is the main cause of poor signal quality.

[0187] Execution: The linear resonant actuator (LRA) of the control device emits a brief, gentle oscillation (e.g., lasting 100 seconds). ms ).

[0188] Design Intent: Vibration is a non-invasive reminder that requires no visual attention, making it suitable for prompting users to "pause their movements and keep their arms still if measurements are needed" in scenarios such as walking or doing housework.

[0189] Visual cues: Triggering scenario: The signal is rejected, but the user is stationary. This suggests that poor signal quality may be due to improper device fit (such as being too loose or misaligned) or other non-motion-related factors.

[0190] Execution: Display a preset icon or short text prompt on the device's screen (if applicable) or the paired mobile app interface, such as "Please adjust the watch's wearing position" or "Please ensure the sensor is in contact with your skin".

[0191] Design intent: To provide more specific and actionable guidance when a user intends to take a measurement (in a static state) but the measurement fails.

[0192] None (No feedback): Triggering scenario: The signal is accepted (regardless of whether it carries a warning). The system has successfully acquired valid data and should not disturb the user.

[0193] Anti-harassment mechanism: To prevent frequent triggering of prompts under continuously harsh environments (such as strenuous exercise), the system has a built-in silent period. After issuing a Vibrate or Visual feedback, the system enters a silent period of [duration missing]. A silence period of 30 seconds (e.g.) will be enforced. During this silence period, even if a new window is decided as Reject, it will be forced... .

[0194] Furthermore, decision state recording and output: This sub-step completes the final closed loop of the current window processing, records the state, and outputs it in a formatted manner.

[0195] Internal status update: The system maintains a fixed-length decision history buffer. Used to record recent W The decision results of each window (e.g.) After processing one window... ,implement: (54), If the buffer is full, the oldest record is removed. This buffer can be used for simple performance monitoring (such as calculating recent signal adoption rates) or for potential future algorithm enhancements.

[0196] Final output encapsulation: The final output of step S500 is a structured instruction tuple that directly drives the hardware and upper-layer applications: (55), in, It is output separately as the raw confidence level for this decision, which can be used by advanced applications (such as health data analytics platforms) for more refined trend analysis or risk assessment.

[0197] The only input for step S500 It relies entirely on the output of the lightweight CNN classifier in step S400. The accuracy of the classification in step S400 directly determines the correctness of the decision in step S500.

[0198] The decision result of step S500 (especially Reject) determines whether the raw PPG data flows to the blood pressure calculation core algorithm. This constitutes the core of the entire "signal quality gating" mechanism, ensuring that the backend core algorithm only processes reliable data, thereby guaranteeing the overall reliability of blood pressure measurement at the system level.

[0199] In summary, the entire process for detecting motion artifacts in blood pressure signals, from multimodal signal acquisition to intelligent decision-making and feedback, has been completed. This process integrates differential and integral feature extraction, quantum feature enhancement, deep learning classification, and rule-based decision feedback, forming a solution that can be efficiently deployed on embedded devices.

[0200] On the other hand, the present invention provides a blood pressure signal motion artifact detection device 200, such as... Figure 9 As shown, in the blood pressure signal motion artifact detection method according to any of the above embodiments, the apparatus includes: The signal acquisition and preprocessing module 210 is used to simultaneously acquire PPG and motion reference signals, and output multi-channel signal segments to be detected after preprocessing and window segmentation. Feature construction module 220 is used to construct static feature vectors and temporal feature matrices based on the multi-channel signal segments to be detected; The quantum feature extraction module 230 is configured to input static feature vectors and output quantum-enhanced feature vectors; The feature fusion classification module 240 is used for feature fusion and outputs quality level probability prediction results based on the trained convolutional neural network classifier. The decision feedback module 250 is used to perform hierarchical decision-making based on the probability prediction results. The hierarchical decision-making includes at least adopting the PPG signal of the current window for blood pressure calculation, adopting it after marking it for blood pressure calculation, or rejecting it for blood pressure calculation, and triggering user feedback when rejecting it.

[0201] Another aspect of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement any of the steps described in the above method embodiments and achieve the same or corresponding technical effects. To avoid repetition, it will not be described again here.

[0202] The blood pressure signal motion artifact detection device provided in this embodiment of the invention can realize the various processes in the above method embodiments, and will not be repeated here to avoid repetition.

[0203] Please see Figure 10 In another aspect, the present invention provides an electronic device 300, a readable storage medium, and a computer program product.

[0204] Figure 10 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present disclosure is shown. This electronic device can be a variety of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers; it can also be a variety of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components, connections, and functions shown herein are merely examples and are not intended to limit the implementation of this application.

[0205] like Figure 10 As shown, device 300 includes a computing unit 310. The computing unit 310 can perform corresponding calculations and processing based on a computer program stored in read-only memory (ROM) 320, or based on a computer program loaded from storage unit 380 into random access memory (RAM) 330. RAM 330 can also store various programs and data required for the operation of device 300. The computing unit 310, ROM 320, and RAM 330 are interconnected via bus 340. Input / output (I / O) interface 350 is also connected to bus 340.

[0206] Multiple components in device 300 are connected to I / O interface 350, including: input unit 360, such as keyboard and mouse; output unit 370, such as various displays and speakers; storage unit 380, such as disk and optical disk; and communication unit 390, such as network card, modem, and wireless transceiver. The communication unit 390 enables device 300 to interact with other devices via the Internet and various telecommunications networks.

[0207] The computing unit 310 can be a general-purpose or special-purpose processing component with computing capabilities. Exemplarily, the computing unit 310 includes, but is not limited to, a central processing unit (CPU), a graphics processing unit (GPU), a dedicated artificial intelligence computing chip, a computing unit running machine learning model algorithms, a digital signal processor (DSP), a processor, a controller, a microcontroller, etc. The computing unit 310 is used to execute the method steps described in the foregoing embodiments of this application, such as a variable quantum circuit training method and a blood pressure signal motion artifact detection method. In some embodiments, the above methods can be implemented as computer software programs and stored in a machine-readable medium such as a storage unit 380. In some embodiments, the computer program can be loaded into and installed into the device 300 via a ROM 320 and / or a communication unit 390. When the computer program is loaded into the RAM 330 and run by the computing unit 310, all or part of the steps of the aforementioned methods can be executed. In other embodiments, the computing unit 310 can also be configured to execute the aforementioned methods via firmware or other means.

[0208] The systems and technical solutions disclosed in this application can be implemented using digital electronic circuits, integrated circuits, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, or combinations thereof. Various implementations can be embodied as one or more computer programs, which can run on a programmable system with at least one programmable processor; the programmable processor can be a general-purpose or special-purpose processor, capable of receiving data and instructions from a storage system and at least one input device, and transmitting data and instructions to the storage system and at least one output device.

[0209] The program code used to implement the methods disclosed herein may be written in one or more programming languages. The program code may be provided to the processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, so that when the processor or controller executes the program code, it implements the functions and operations defined in the flowcharts and block diagrams. The program code may run entirely on a local device, partially on a local device, or as an independent software package, partially on a local device and partially on a remote device, or entirely on a remote device or server.

[0210] In the context of this disclosure, a machine-readable medium is a tangible medium that stores programs for use by an instruction execution system, apparatus, or device. Machine-readable storage media include, but are not limited to: electrical circuits, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM, flash memory), optical fibers, portable optical disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, and any combination of the above media.

[0211] To achieve human-computer interaction, the system and technical solution of this application can be deployed on a computer equipped with a display device, keyboard, and positioning device; the display device is such as a cathode ray tube (CRT) monitor or a liquid crystal display (LCD), and the positioning device is such as a mouse or trackball. Users can input commands to the computer through the keyboard and positioning device. Other devices can also be used to achieve human-computer interaction, such as providing feedback to users through visual, auditory, or tactile means, and receiving user input through voice or touch.

[0212] The system and technical solution of this application can be deployed on a data server containing backend components, an application server containing middleware components, a user terminal containing frontend components, or a combination of the above components. The various components of the system can be connected via various digital communication networks such as local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

[0213] The computer system may include clients and servers, which typically interact remotely via a communication network, forming a client-server architecture through corresponding computer programs. The server may be a cloud server, a distributed server, or a server incorporating blockchain technology.

[0214] It should be understood that the execution order of the method steps described in this application may be adjusted, or steps may be added or deleted, as long as the expected technical effect of this application can be achieved, and this application does not limit this.

[0215] The specific embodiments described above are not intended to limit the scope of protection of this application. Those skilled in the art can modify, combine, split, or replace this application according to design requirements; all modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A variational quantum circuit training method for providing a quantum-enhanced feature extraction module for blood pressure signal motion artifact detection, characterized in that, Includes the following steps: Obtain a training dataset containing multiple multi-channel signal segments. Each multi-channel signal segment is obtained by preprocessing and window segmentation of the synchronously acquired raw PPG signal and raw motion reference signal, and each multi-channel signal segment corresponds to a real label. Construct the static feature vector and temporal feature matrix for each of the multi-channel signal segments; The static feature vector is input into the quantum feature extraction module for quantum feature extraction processing to obtain the quantum enhanced feature vector; The quantum-enhanced feature vector and the temporal feature matrix are fused together, and the fused features are input into a convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. Based on the change intensity of the quantum enhancement eigenvector corresponding to adjacent time windows and the standard deviation of the original PPG signal differential sequence within the corresponding window, the differential consistency constraint loss is calculated; wherein, the differential consistency constraint loss is used to ensure that the transformation characteristics of the quantum enhancement eigenvector relative to the static eigenvector are consistent with the differential characteristics of the original PPG signal and the original motion reference signal. The classification loss is calculated based on the probability prediction results and the true labels, and the classification loss is weighted and summed with the differential consistency constraint loss to obtain the total loss. With minimizing the overall loss as the optimization objective, the trainable parameters of the quantum feature extraction module and the convolutional neural network classifier are iteratively updated using the gradient descent method to complete the training.

2. The method of claim 1, wherein, In the step of obtaining the training dataset, the preprocessing and window segmentation include the following steps: The original PPG signal and the original motion reference signal were synchronously sampled using the same master clock, and the sampling frequency was set to be no less than 100 Hz. The original PPG signal is subjected to a first bandpass filter with a passband frequency range of 0.5 Hz to 8.0 Hz; The original motion reference signal is subjected to a second bandpass filter with a passband frequency range of 0.1 Hz to 20.0 Hz; The filtered PPG signal and the motion reference signal are windowed to obtain the original multi-channel signal segments. The original multi-channel signal segment is normalized to eliminate signal amplitude scaling and DC offset caused by individual and equipment differences, thus obtaining the multi-channel signal segment.

3. The method of claim 1, wherein, Constructing the static feature vector and temporal feature matrix of the multi-channel signal segment includes the following steps: The first-order difference and the second-order difference are calculated for the PPG signal and the motion reference signal in the multi-channel signal segment, respectively, to obtain the differential sequence. Based on the differential sequence, a first set of scalar features is extracted. The first set of scalar features includes at least one of standard deviation, kurtosis and zero-crossing rate. The absolute integral and square integral within the sliding window are calculated for the PPG signal and motion reference signal in the multi-channel signal segment, respectively, to obtain an integral sequence. Based on the integral sequence, a second set of scalar features is extracted. The second set of scalar features includes at least one of mean, coefficient of variation, and root mean square. Calculate the normalized cross-correlation function between the differential sequence of the PPG signal and the differential sequence of the motion reference signal, extract the maximum value of the cross-correlation coefficient and its corresponding time delay, and calculate the energy ratio of the PPG signal and the motion reference signal, which are used as the third set of scalar features. The first set of scalar features, the second set of scalar features, and the third set of scalar features are concatenated into a static feature vector, and one or more of the original PPG signal, the original motion reference signal, and its first-order differential sequence are retained and arranged in time sequence to form a time-series feature matrix.

4. The method of claim 1, wherein, The quantum feature extraction module includes a classical encoder, an angle encoding layer, a variable quantum circuit, a measurement layer, and a linear layer.

5. The method according to claim 4, characterized in that, The step of inputting the static feature vector into the quantum feature extraction module for quantum feature extraction processing to obtain the quantum enhanced feature vector includes: The static feature vector is mapped to a rotation angle parameter using the classic encoder. The rotation angle parameter is loaded onto the rotation gate of the qubit through the angle encoding layer to generate the input quantum state of the quantum circuit. The input quantum state is transformed using the parametric quantum gates and entanglement gates in the variable quantum circuit to obtain the final quantum state; The final quantum state is measured through the measurement layer, and the measurement result is linearly transformed through the linear layer to obtain the quantum enhancement feature vector.

6. The method according to claim 1, characterized in that, The feature fusion of the quantum-enhanced feature vector and the temporal feature matrix includes: The quantum enhancement feature vector is copied along the time dimension to generate a quantum feature matrix with the same time length and identical rows as the temporal feature matrix; The temporal feature matrix and the quantum feature matrix are concatenated along the feature channel dimension to obtain a fused feature matrix; The fused feature matrix is ​​converted into a three-dimensional tensor.

7. The method according to claim 1, characterized in that, The convolutional neural network consists of a depthwise separable convolutional layer, a dilated convolutional layer, a global average pooling layer, and at least one fully connected layer connected in series.

8. The method according to claim 1, characterized in that, The gradient descent method uses the Adam optimizer to perform parameter iteration; for the trainable parameters of the variable quantum circuit, the gradient is calculated using parameter shift rules or automatic differentiation.

9. A variable quantum circuit training apparatus for implementing the variable quantum circuit training method according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire a training dataset, which contains multiple multi-channel signal segments. Each multi-channel signal segment is obtained by preprocessing and window segmentation of the synchronously acquired raw PPG signal and raw motion reference signal, and each multi-channel signal segment corresponds to a real label. The feature construction module is used to construct the static feature vector and temporal feature matrix for each of the multi-channel signal segments; A quantum feature extraction module is used to perform quantum feature extraction processing on the static feature vector to obtain a quantum enhanced feature vector; The fusion classification module is used to fuse the quantum-enhanced feature vector with the temporal feature matrix, and input the fused features into a convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. The loss calculation module is used to calculate the differential consistency constraint loss based on the change intensity of the quantum enhancement feature vector corresponding to adjacent time windows and the standard deviation of the original PPG signal difference sequence within the corresponding window, calculate the classification loss based on the probability prediction result and the true label, and perform a weighted summation of the classification loss and the differential consistency constraint loss to obtain the total loss. The parameter update module is used to iteratively update the trainable parameters of the quantum feature extraction module and the convolutional neural network classifier using gradient descent, with the goal of minimizing the overall loss, to complete the training.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the variable quantum circuit training method as described in any one of claims 1 to 8.

11. A method for detecting motion artifacts in blood pressure signals, characterized in that, The quantum feature extraction module and convolutional neural network classifier trained based on the training method described in any one of claims 1 to 8 include the following steps: The original PPG signal and the original motion reference signal are acquired synchronously, preprocessed and windowed to obtain the multi-channel signal segment to be detected; Construct the static feature vector and temporal feature matrix of the multi-channel signal segment to be detected; The static feature vector is input into the trained quantum feature extraction module to obtain the quantum enhanced feature vector; The quantum-enhanced feature vector is fused with the temporal feature matrix, and the fused features are input into the trained convolutional neural network classifier to obtain the probability prediction result of the PPG signal quality level under the current time window. Based on the probability prediction results, a hierarchical decision is executed. The hierarchical decision includes at least adopting the PPG signal of the current window for blood pressure calculation, adopting it after marking it for blood pressure calculation, or rejecting it for blood pressure calculation, and triggering user feedback when rejecting it.

12. The method according to claim 11, characterized in that, The probability prediction results include a first quality level, a second quality level, and a third quality level, and the step of performing a graded decision based on the probability prediction results includes: Set a high confidence threshold and a low confidence threshold, wherein the high confidence threshold is greater than the low confidence threshold; The category corresponding to the maximum value among the probabilities of the first quality level, the second quality level, and the third quality level is taken as the predicted main category. When the predicted main category is the first quality level and its probability is greater than or equal to the high confidence threshold, the decision is to adopt the current signal. When the predicted main category is the first quality level and its probability is between the low confidence threshold and the high confidence threshold, or when the predicted main category is the second quality level, or when the predicted main category is the third quality level and its probability is less than the low confidence threshold, the decision is to label and then adopt. When the predicted main category is the third quality level and its probability is greater than or equal to the high confidence threshold, the decision is to reject the current signal.

13. The method according to claim 11, characterized in that, When the decision is to adopt the current signal, the PPG signal of the current window is passed to the blood pressure calculation algorithm, and the calculated blood pressure value is marked as a high-confidence measurement result.

14. The method according to claim 11, characterized in that, When the decision is to adopt the marked value, the PPG signal of the current window is passed to the blood pressure calculation algorithm, and the calculated blood pressure value is marked as a low-confidence measurement result.

15. The method according to claim 11, characterized in that, When the decision is to reject the current signal, discard the PPG signal of the current window, do not trigger the blood pressure calculation process, and start the user feedback mechanism.

16. The method according to claim 15, characterized in that, The triggering of user feedback includes: Obtain the user's current state, which includes both active and inactive states; When the decision is to reject the current signal and the user is active, a vibration alert is triggered. A visual cue is triggered when the decision is to reject the current signal and the user is stationary.

17. The method according to claim 16, characterized in that, Also includes: After a vibration or visual cue is triggered, the system enters a preset silent period during which user feedback is prohibited from being triggered again.

18. A device for detecting motion artifacts in blood pressure signals, characterized in that, Based on the blood pressure signal motion artifact detection method according to any one of claims 11 to 17, the apparatus comprises: The signal acquisition and preprocessing module is used to simultaneously acquire PPG and motion reference signals, and output multi-channel signal segments to be detected after preprocessing and window segmentation. The feature construction module is used to construct static feature vectors and temporal feature matrices based on the multi-channel signal segments to be detected. The quantum feature extraction module is configured to take a static feature vector as input and output a quantum-enhanced feature vector. The feature fusion classification module is used to fuse features and output quality level probability prediction results based on the trained convolutional neural network classifier. The decision feedback module is used to perform hierarchical decisions based on probability prediction results. The hierarchical decisions include at least adopting the PPG signal of the current window for blood pressure calculation, adopting it after marking it for blood pressure calculation, or rejecting it for blood pressure calculation, and triggering user feedback when rejecting it.

19. An electronic device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the blood pressure signal motion artifact detection method as described in any one of claims 11 to 17.

20. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the steps of the blood pressure signal motion artifact detection method as described in any one of claims 11 to 17.

Images